Re: [PATCH net-next 00/12] Unmask upper DSCP bits - part 2

[Guillaume Nault <gnault@redhat.com>]

On Wed, Aug 28, 2024 at 03:09:19PM +0300, Ido Schimmel wrote:
> On Tue, Aug 27, 2024 at 06:45:53PM +0300, Ido Schimmel wrote:
> > On Tue, Aug 27, 2024 at 03:47:05PM +0200, Guillaume Nault wrote:
> > > On Tue, Aug 27, 2024 at 02:18:01PM +0300, Ido Schimmel wrote:
> > > > tl;dr - This patchset continues to unmask the upper DSCP bits in the
> > > > IPv4 flow key in preparation for allowing IPv4 FIB rules to match on
> > > > DSCP. No functional changes are expected. Part 1 was merged in commit
> > > > ("Merge branch 'unmask-upper-dscp-bits-part-1'").
> > > > 
> > > > The TOS field in the IPv4 flow key ('flowi4_tos') is used during FIB
> > > > lookup to match against the TOS selector in FIB rules and routes.
> > > > 
> > > > It is currently impossible for user space to configure FIB rules that
> > > > match on the DSCP value as the upper DSCP bits are either masked in the
> > > > various call sites that initialize the IPv4 flow key or along the path
> > > > to the FIB core.
> > > > 
> > > > In preparation for adding a DSCP selector to IPv4 and IPv6 FIB rules, we
> > > 
> > > Hum, do you plan to add a DSCP selector for IPv6? That shouldn't be
> > > necessary as IPv6 already takes all the DSCP bits into account. Also we
> > > don't need to keep any compatibility with the legacy TOS interpretation,
> > > as it has never been defined nor used in IPv6.
> > 
> > Yes. I want to add the DSCP selector for both families so that user
> > space would not need to use different selectors for different families.
> 
> Another approach could be to add a mask to the existing tos/dsfield. For
> example:
> 
> # ip -4 rule add dsfield 0x04/0xfc table 100
> # ip -6 rule add dsfield 0xf8/0xfc table 100
> 
> The default IPv4 mask (when user doesn't specify one) would be 0x1c and
> the default IPv6 mask would be 0xfc.
> 
> WDYT?

For internal implementation, I find the mask option elegant (to avoid
conditionals). But I don't really like the idea of letting user space
provide its own mask. This would let the user create non-standard
behaviours, likely by mistake (as nobody seem to ever have requested
that flexibility).

I think my favourite approach would be to have the new FRA_DSCP
attribute work identically on both IPv4 and IPv6 FIB rules and keep
the behaviour of the old "tos" field of struct fib_rule_hdr unchanged.

This "tos" field would still work differently for IPv4 and IPv6, as it
always did, but people wanting consistent behaviour could just use
FRA_DSCP instead. Also, FRA_DSCP accepts real DSCP values as defined in
RFCs, while "tos" requires the 2 bits shift. For all these reasons, I'm
tempted to just consider "tos" as a legacy option used only for
backward compatibility, while FRA_DSCP would be the "clean" interface.

Is that approach acceptable for you?